[PATCH] m32r: update switch_to macro for tuning
authorHirokazu Takata <takata@linux-m32r.org>
Wed, 19 Apr 2006 05:21:38 +0000 (22:21 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 19 Apr 2006 16:13:51 +0000 (09:13 -0700)
- Remove unnecessary push/pop's of the switch_to() macro
  for performance tuning.
- Cosmetic updates: change __inline__ to inline, etc.

Signed-off-by: Hirokazu Takata <takata@linux-m32r.org>
Cc: NIIBE Yutaka <gniibe@fsij.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
arch/m32r/kernel/entry.S
include/asm-m32r/system.h

index 5e4a0c8a5d3c976729b68de3d527e20c4f6ef2f8..920bb742b7a2a90394c133998809917f0bb33f56 100644 (file)
@@ -132,7 +132,7 @@ VM_MASK             = 0x00020000
 #endif
 
 ENTRY(ret_from_fork)
-       ld      r0, @sp+
+       pop     r0
        bl      schedule_tail
        GET_THREAD_INFO(r8)
        bra     syscall_exit
@@ -310,7 +310,7 @@ ENTRY(ei_handler)
 ;    GET_ICU_STATUS;
        seth    r0, #shigh(M32R_ICU_ISTS_ADDR)
        ld      r0, @(low(M32R_ICU_ISTS_ADDR),r0)
-       st      r0, @-sp
+       push    r0
 #if defined(CONFIG_SMP)
        /*
         * If IRQ == 0      --> Nothing to do,  Not write IMASK
@@ -547,7 +547,7 @@ check_end:
 #endif  /* CONFIG_PLAT_M32104UT */
        bl      do_IRQ
 #endif  /* CONFIG_SMP */
-       ld      r14, @sp+
+       pop     r14
        seth    r0, #shigh(M32R_ICU_IMASK_ADDR)
        st      r14, @(low(M32R_ICU_IMASK_ADDR),r0)
 #else
index c5ab5da56d21acb42dfb4c18e5bbdd8151aa58f8..e55013f378e582357df670aee5638e62e8e5a2af 100644 (file)
@@ -6,8 +6,8 @@
  * License.  See the file "COPYING" in the main directory of this archive
  * for more details.
  *
- * Copyright (C) 2001  by Hiroyuki Kondo, Hirokazu Takata, and Hitoshi Yamamoto
- * Copyright (C) 2004  Hirokazu Takata <takata at linux-m32r.org>
+ * Copyright (C) 2001  Hiroyuki Kondo, Hirokazu Takata, and Hitoshi Yamamoto
+ * Copyright (C) 2004, 2006  Hirokazu Takata <takata at linux-m32r.org>
  */
 
 #include <linux/config.h>
  * switch_to(prev, next) should switch from task `prev' to `next'
  * `prev' will never be the same as `next'.
  *
- * `next' and `prev' should be struct task_struct, but it isn't always defined
+ * `next' and `prev' should be task_t, but it isn't always defined
  */
 
 #define switch_to(prev, next, last)  do { \
-       register unsigned long  arg0 __asm__ ("r0") = (unsigned long)prev; \
-       register unsigned long  arg1 __asm__ ("r1") = (unsigned long)next; \
-       register unsigned long  *oldsp __asm__ ("r2") = &(prev->thread.sp); \
-       register unsigned long  *newsp __asm__ ("r3") = &(next->thread.sp); \
-       register unsigned long  *oldlr __asm__ ("r4") = &(prev->thread.lr); \
-       register unsigned long  *newlr __asm__ ("r5") = &(next->thread.lr); \
-       register struct task_struct  *__last __asm__ ("r6"); \
        __asm__ __volatile__ ( \
-               "st     r8, @-r15                                 \n\t" \
-               "st     r9, @-r15                                 \n\t" \
-               "st    r10, @-r15                                 \n\t" \
-               "st    r11, @-r15                                 \n\t" \
-               "st    r12, @-r15                                 \n\t" \
-               "st    r13, @-r15                                 \n\t" \
-               "st    r14, @-r15                                 \n\t" \
-               "seth  r14, #high(1f)                             \n\t" \
-               "or3   r14, r14, #low(1f)                         \n\t" \
-               "st    r14, @r4    ; store old LR                 \n\t" \
-               "st    r15, @r2    ; store old SP                 \n\t" \
-               "ld    r15, @r3    ; load new SP                  \n\t" \
-               "st     r0, @-r15  ; store 'prev' onto new stack  \n\t" \
-               "ld    r14, @r5    ; load new LR                  \n\t" \
-               "jmp   r14                                        \n\t" \
-               ".fillinsn                                        \n  " \
-               "1:                                               \n\t" \
-               "ld     r6, @r15+  ; load 'prev' from new stack   \n\t" \
-               "ld    r14, @r15+                                 \n\t" \
-               "ld    r13, @r15+                                 \n\t" \
-               "ld    r12, @r15+                                 \n\t" \
-               "ld    r11, @r15+                                 \n\t" \
-               "ld    r10, @r15+                                 \n\t" \
-               "ld     r9, @r15+                                 \n\t" \
-               "ld     r8, @r15+                                 \n\t" \
-               : "=&r" (__last) \
-               : "r" (arg0), "r" (arg1), "r" (oldsp), "r" (newsp), \
-                 "r" (oldlr), "r" (newlr) \
-               : "memory" \
+               "       seth    lr, #high(1f)                           \n" \
+               "       or3     lr, lr, #low(1f)                        \n" \
+               "       st      lr, @%4  ; store old LR                 \n" \
+               "       ld      lr, @%5  ; load new LR                  \n" \
+               "       st      sp, @%2  ; store old SP                 \n" \
+               "       ld      sp, @%3  ; load new SP                  \n" \
+               "       push    %1  ; store `prev' on new stack         \n" \
+               "       jmp     lr                                      \n" \
+               "       .fillinsn                                       \n" \
+               "1:                                                     \n" \
+               "       pop     %0  ; restore `__last' from new stack   \n" \
+               : "=r" (last) \
+               : "0" (prev), \
+                 "r" (&(prev->thread.sp)), "r" (&(next->thread.sp)), \
+                 "r" (&(prev->thread.lr)), "r" (&(next->thread.lr)) \
+               : "memory", "lr" \
        ); \
-       last = __last; \
 } while(0)
 
 /*
@@ -167,8 +146,8 @@ extern void  __xchg_called_with_bad_pointer(void);
 #define DCACHE_CLEAR(reg0, reg1, addr)
 #endif /* CONFIG_CHIP_M32700_TS1 */
 
-static __inline__ unsigned long __xchg(unsigned long x, volatile void * ptr,
-       int size)
+static inline unsigned long
+__xchg(unsigned long x, volatile void * ptr, int size)
 {
        unsigned long flags;
        unsigned long tmp = 0;
@@ -220,7 +199,7 @@ static __inline__ unsigned long __xchg(unsigned long x, volatile void * ptr,
 
 #define __HAVE_ARCH_CMPXCHG    1
 
-static __inline__ unsigned long
+static inline unsigned long
 __cmpxchg_u32(volatile unsigned int *p, unsigned int old, unsigned int new)
 {
        unsigned long flags;
@@ -254,7 +233,7 @@ __cmpxchg_u32(volatile unsigned int *p, unsigned int old, unsigned int new)
    if something tries to do an invalid cmpxchg().  */
 extern void __cmpxchg_called_with_bad_pointer(void);
 
-static __inline__ unsigned long
+static inline unsigned long
 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
 {
        switch (size) {